Skip to main content

DeepSeek's New Flash Model Lands on Qwen, and It's Cheap

DeepSeek's New Flash Model Lands on Qwen, and It's Cheap

If you've been hunting for a powerful yet affordable AI model, DeepSeek's latest release might just be your answer. DeepSeek-V4.1-Flash is now live on the Qwen AI platform, and it's bringing some serious specs to the table.

What's Under the Hood?

This isn't just another incremental update. The V4.1-Flash is DeepSeek's new lightweight flagship, built on a Mixture-of-Experts (MoE) architecture with a whopping 552 billion total parameters. But here's the clever part: it only activates about 8 billion parameters for input and 16 billion for output. That means you get the muscle of a giant model without the usual computational heft.

It also handles both text and images natively, and the context window? A massive 1 million tokens—enough to digest entire books or sprawling codebases in one go. Output can stretch to roughly 393,000 tokens, making it a beast for long-form generation.

Image

The Cost Story

Let's talk money, because that's where things get interesting. DeepSeek has squeezed the KV Cache requirements down to a quarter of the previous generation for HBM and an eighth for SSD storage. Translation: long contexts and multi-turn tool calls won't drain your resources as quickly.

On the Qwen platform, pricing is time-based. During off-peak hours, you'll pay 1 yuan per million input tokens and 4 yuan per million output tokens. Peak hours double that to 2 yuan and 8 yuan, respectively. For small teams, that off-peak rate is a steal—especially combined with the Token Plan subscription, which lets you fire off batch inferences without breaking the bank.

Ready for Deployment

Alibaba Cloud's BaiLian has already worked with the vLLM community to get everything adapted. You can access the model on both the China site and international sites, covering regions from Beijing to Singapore, the US, Germany, Japan, and Hong Kong. It supports multi-turn conversations, function calls, online search, context caching, and structured output. The max_tokens limit sits at around 393,216.

For enterprises, this means you can migrate tasks like long document parsing, customer service knowledge bases, code repository Q&A, and multimodal order review to a single interface. No more juggling multiple models for different jobs.

Why It Matters

Industry watchers say that by folding this "large parameters, small activation, strong caching" combo into the Qwen ecosystem, V4.1-Flash slashes the trial-and-error costs of building long-context agents. And for smaller teams, the low off-peak pricing plus the Token Plan subscription opens the door to flexible batch inference and rapid prototyping.

So if you've been waiting for a model that balances power, speed, and price, this might be the moment to dive in.

Key Points

  • DeepSeek-V4.1-Flash launches on Qwen AI platform with API and Token Plan access.
  • MoE architecture with 552B total parameters, but only ~8B input and ~16B output activated.
  • Supports text and images, with a 1M token context and ~393K token output.
  • Aggressive pricing: 1 yuan/million input tokens off-peak, 4 yuan/million output tokens.
  • Available on Alibaba Cloud BaiLian across multiple global regions, ready for enterprise tasks.